Predictability of Hurricane Storm Surge: An Ensemble Forecasting Approach Using Global Atmospheric Model Data

Abstract

Providing storm surge risk information at multi-day lead times is critical for hurricane evacuation decisions, but predictability of storm surge inundation at these lead times is limited. This study develops a method to parameterize and adjust tropical cyclones derived from global atmospheric model data, for use in storm surge research and prediction. We implement the method to generate storm tide (surge + tide) ensemble forecasts for Hurricane Michael (2018) at five initialization times, using archived operational ECMWF ensemble forecasts and the dynamical storm surge model ADCIRC. The results elucidate the potential for extending hurricane storm surge prediction to several-day lead times, along with the challenges of predicting the details of storm surge inundation even 18 h before landfall. They also indicate that accurately predicting Hurricane Michael’s rapid intensification was not needed to predict the storm surge risk. In addition, the analysis illustrates how this approach can help identify situationally and physically realistic scenarios that pose greater storm surge risk. From a practical perspective, the study suggests potential approaches for improving real-time probabilistic storm surge prediction. The method can also be useful for other applications of atmospheric model data in storm surge research, forecasting, and risk analysis, across weather and climate time scales.

1. Introduction

Storm surges generated by tropical cyclones (TCs) are major threats to lives and property worldwide—including in the U.S., as Hurricane Ian recently demonstrated. Motivated in part by devastating TC storm surge impacts in the 2000s [1,2], the research community and U.S. National Weather Service (NWS) have advanced multiple aspects of storm surge prediction and risk communication over the last two decades [3,4,5,6,7,8,9]. However, for many U.S. coastal hurricane evacuation decisions, storm surge inundation forecasts are needed at longer lead times than they are currently available [10,11,12,13,14]. As sea levels rise due to climate change and coastal populations increase, the importance of storm surge forecast information at several-day lead times is likely to continue growing.

TC-induced storm surge inundation is challenging to forecast for several reasons, including the dependence of location-specific water levels on details of a TC’s track, intensity, and structure [3,6,15,16,17,18,19]. Thus, although many storm surge studies focus on the accuracy of deterministic (single-valued) simulations and predictions, the NWS National Hurricane Center (NHC) and some other prediction centers forecast storm surge using an ensemble approach [20]. For example, NHC’s operational storm surge prediction system generates probabilistic forecasts using a computationally efficient storm surge model, SLOSH (Sea Lakes and Overland Surges from Hurricanes), which enables running many simulations. SLOSH is forced with parameterized vortices defined by TC track, intensity, radius of maximum wind, and other parameters, and storm surge ensembles are generated by systematically perturbing TC parameters and rerunning SLOSH for each perturbed TC [3,6,21].

This study uses an ensemble storm surge modeling approach to investigate the predictability of TC-induced storm surge at multiple forecast lead times. More specifically, we examine predictability of coastal inundation due to storm-induced surge combined with tides. Although the technical terminology for this combined hazard is “storm tide”, we follow NHC and many others by referring to this as “storm surge” or “surge”. This is consistent with the study’s focus on predictability: since tides are highly predictable, the hazard’s limited predictability arises primarily from the storm-induced surge. The study does not include other factors that can contribute to storm-induced water levels, such as waves and precipitation runoff, but the approach could readily be adapted to use techniques for coastal water level simulation that incorporate those factors.

Storm surge ensemble approaches that start with a control simulation and independently perturb TC parameters, similar to NHC’s current method, have proven useful for both research and real-time prediction [3,17,21,22,23,24,25,26,27]. However, the resulting storm surge ensembles do not reflect the integrated, multi-parameter atmospheric uncertainty associated with a particular TC. Another approach to generating storm surge ensembles is using gridded output from an atmospheric numerical model ensemble to force a set of storm surge model simulations. Global atmospheric model simulations from forecast centers can be used for this type of work, but they may lack the spatial resolution, or other features required to adequately simulate storm surge magnitude, especially for strong or small TCs [28,29,30,31]. Consequently, this type of work often uses a higher-resolution, regional atmospheric model ensemble as the meteorological forcing [15,16,32,33,34]. Such simulations are usually run retrospectively rather than in real time due to the logistics and computational expense involved.

Here, we employ an intermediate approach that derives TC tracks and other parameters from a global numerical weather prediction ensemble, and then uses the resulting ensemble of parameterized TCs to force a dynamical storm surge model. This approach allows us to utilize operational weather prediction ensemble data, which are readily available for a variety of TCs at a range of lead times, to generate ensembles of parameterized TC forecasts that are consistent with realistic, storm-specific atmospheric scenarios. Using parameterized TCs also facilitates adjusting the ensemble for biases in the atmospheric model data. We use global weather prediction model data because doing so enables extending the research to longer forecast lead times. Although specifying TCs with a small number of parameters alters details of the storm structure, parameterized TCs are frequently used in storm surge simulation and prediction [18,23,26,35,36,37], and this method is appropriate for our ensemble approach and study goals. Further, parameterizing TCs from available atmospheric model data is more computationally efficient than running an ensemble of high-resolution atmospheric simulations, and thus more practical to implement in real time or for a range of cases.

The article focuses on Hurricane Michael, which initially formed as a tropical depression in the Gulf of Mexico at approximately 6 UTC 7 October 2018, and then rapidly organized and intensified. The storm made landfall in the Florida panhandle 3.5 days later as a Category 5 hurricane (Figure 1), causing extensive damage from both wind and storm surge [38]. Hurricane Michael presents an interesting case for storm surge prediction for several reasons, including its rapid intensification, the associated challenges of accurately forecasting its intensity evolution, and its formation only a few days prior to landfall [38,39,40,41,42]. Although the article examines one TC, the type of atmospheric forcing data used is readily available, and so the methods can be extended to investigate other storms.

The research has three goals:

• Develop and test a method for parameterizing tropical cyclones from atmospheric model simulations for use in storm surge simulation, prediction, and research.
• Use the method, in conjunction with numerical weather prediction ensemble data and a dynamical storm surge model, to generate ensemble TC–storm surge predictions for Hurricane Michael.
• Analyze the resulting ensemble data to:
  • investigate the predictability of storm surge inundation across a range of forecast lead times, from approximately 1–3 days, and
  • explore the importance of different TC parameters for predicting storm surge inundation at different lead times.

The method in #1, combined with the approach in #2, enables us to generate storm surge ensembles that simulate uncertainties in inundation associated with the TC forecast uncertainties represented in atmospheric ensembles. Here we use the methods for storm surge predictability research, but they are also applicable for other types of work, including practical storm surge prediction and longer-term storm surge risk analyses using climate or weather model data.

We use atmospheric data from the ECMWF (European Centre for Medium-Range Weather Forecasts) Ensemble Prediction System [43] and the dynamical storm surge model ADCIRC (ADvanced CIRCulation Model; [44]). However, a similar approach can be implemented with other types of atmospheric model data and storm surge forecasting platforms. This includes other dynamical storm surge hazard models, as well as potentially machine learning or surrogate model techniques for storm surge prediction [7,27,45,46]. The study also illustrates an approach for adjusting the evolution of TCs in atmospheric model simulations for biases, enabling improved representation of TC and storm surge forecast uncertainties. Here, we adjust storm intensities, but the approach could be adapted to adjust other aspects of TC forecasts, such as storm track or forward speed.

To analyze storm surge predictability at multiple forecast lead times, we generated TC–storm surge ensemble forecasts at five initialization times, every 12 h from 0 UTC 8 October to 0 UTC 10 October 2018 (

Table 1. Summary of peak TC intensities in the NHC OFCL forecast, the parameterized TC ensemble forecasts derived from the ECMWF data, and the intensity-adjusted parameterized TC ensemble forecasts, for each of the five initialization times studied. Vmax1_L and Pmin_L represent the maximum 1 min average wind speed (Vmax1, kts) and minimum central pressure (mb), respectively, over a TC’s lifetime (through 0 UTC 13 October). For the ensembles, the mean and range of Vmax1_L and Pmin_L across the 50 ensemble members are shown. For Best track, Vmax1_L = 140 kt and Pmin_L = 919 mb.

Initialization Time	Approximate Lead Time Prior to Landfall	OFCL	Parameterized TC Ensemble: before Intensity Adjustment		Parameterized TC Ensemble: after Intensity Adjustment
		Vmax1_L	Vmax1_L Ensemble Mean (Range)	Pmin1_L Ensemble Mean (Range)	Vmax1_L Ensemble Mean (Range)	Pmin1_L Ensemble Mean (Range)
0 UTC 8 October	66 h	85	74 (53–96)	965 (945–993)	90 (65–121)	963 (935–983)
12 UTC 8 October	54 h	105	70 (56–85)	970 (951–989)	109 (88–140)	948 (919–963)
0 UTC 9 October	42 h	105	82 (73–94)	956 (937–971)	108 (83–128)	949 (932–968)
12 UTC 9 October	30 h	110	83 (66–99)	956 (944–976)	113 (83–134)	944 (926–968)
0 UTC 10 October	18 h	115	83 (62–97)	958 (947–974)	118 (95–146)	940 (915–959)

, Figure 2). We also ran a storm surge simulation forced by NHC’s Best track data for Hurricane Michael, which is used to evaluate the ensemble forecasts, and a set of experiments with perturbed Best track parameters to illustrate aspects of the study methods (Figure 2). We compare the Best track surge simulation with observations to validate that it is reasonable, but our primary goal is not to develop skillful deterministic storm surge predictions. Instead, we aim to investigate storm surge predictability, while also developing a method that can be applied for a variety of TCs. As a step towards improving storm surge predictions, however, we designed the method to use data that are available for real-time forecasting.

Although the analysis presented focuses on geophysical aspects of TC storm surge, it is informed by the growing body of work on TC hazard communication and decision making [13,47,48,49,50,51,52,53,54,55]. For example, when investigating the predictability of storm surge, we use metrics related to inundation in locations that are normally dry (above mean high tide) because that is the type of information needed for many protective decisions. We also examine inundation timing, since this is important for decision making. Because many decisions rely on location-specific inundation, we investigate the ensembles’ potential to identify locations at greater or lesser storm surge risk. However, given the major challenges of predicting location-specific storm surge at longer lead times, we also explore the use of spatially integrated inundation measures, as a step towards developing metrics that may facilitate providing the longer lead time information needed for decision making given limited predictability [17].

Section 2 describes the study methods and data, including the design and implementation of the TC–storm surge ensemble forecasts that are the focus of the study. This section also describes the Best track storm surge simulation, the Best track experiments, and the metrics used for data analysis. Section 3 presents results from the Best track simulation and experiments, and Section 4 presents results from the TC–storm surge ensemble forecasts. The article closes with a summary and discussion.

2. Materials and Methods

The methods for this study and associated storm surge simulations are summarized in Figure 2. The simulations were run using ADCIRC (v55.02), an open source, shallow water model that incorporates tidal and atmospheric forces [56,57]. All simulations used parameterized TCs generated using the Generalized Asymmetric Holland Model (GAHM; [18]) in ADCIRC (v55.02, option NWS = 20) as the meteorological forcing. Each parameterized TC is defined by its position (latitude and longitude), maximum sustained surface winds (Vmax), minimum sea-level pressure (Pmin), and wind radii representing the maximum extent of sustained 34, 50, and 64 kt winds (R34, R50, and R64, respectively) in 4 quadrants (northeast, southeast, southwest, northwest). Within ADCIRC, the GAHM is used to transform sequences of TC parameters into asymmetric vortex wind and pressure fields, which are then used as atmospheric forcing in ADCIRC’s ocean circulation model. We selected the GAHM because it allows a TC vortex to be fitted to multiple wind radii in each quadrant, which generates more realistic parametric vortices than a symmetric vortex model.

We used parameterized storms because, as discussed in Section 2.1, this enables us to use the ECMWF ensemble forecast data archived at 6 h intervals, and it facilitates adjusting the TC intensity bias in the ECMWF ensemble. Parameterizing a TC simplifies its wind field, which can affect storm surge simulation results; however, this study focuses on ensemble forecasts, rather than the accuracy of individual storm surge simulations. Although ADCIRC includes the capability to simulate waves, e.g., with the SWAN (Simulating WAves Nearshore) model [58], we did not include waves in our study to simplify the methodology for demonstration.

The study used three types of meteorological data: NHC’s Best track for Hurricane Michael, NHC’s official forecasts for the storm (OFCL), and forecast data from the 50-member operational ECMWF ensemble. Best track data represent the actual evolution of a TC in terms of the parameters described above, based on post-storm analysis conducted by NHC [59]. OFCL and ECMWF represent different types of forecast information available as Hurricane Michael approached the U.S. coast; OFCL data are in the form of parameterized TCs similar to Best track, and the ECMWF data used are in the form of spatially gridded meteorological fields. As described in subsequent sections, we ran ADCIRC simulations using two types of meteorological forcing: (1) variations of Best track and (2) parameterized TCs derived from the ECMWF ensemble forecasts, with intensity scaled using OFCL. While (1) must be run retrospectively, (2) requires only data available in near-real time. We also explored using OFCL as the meteorological forcing, but as discussed further in Section 3.2, some of these simulations produced parameterized TC vortices that evolved unrealistically due to the longer (12 or 24 h) forecast intervals in the OFCL data; thus, these simulations are not used in this article.

Implementing the study’s methods involved reconciling the different averaging periods used in wind speed data. The Vmax values in Best track and OFCL represent 1 min average wind speeds (V1), but ADCIRC’s dynamic core is calibrated for 10 min average winds (V10). To convert V1 to V10, the ADCIRC code multiplies parameterized TC Vmax input by a factor one2ten = 0.8928 [58,60]. The gridded ECMWF wind fields, on the other hand, represent V10 [61]. For the ADCIRC meteorological forcing, we kept all ECMWF ensemble wind speeds in terms of V10, as described below. However, to compare Vmax across different types of meteorological data in the figures and tables, we divided ECMWF Vmax10 by one2ten to convert to Vmax1, since that is how Best track and OFCL wind speeds are typically depicted.

The standard format for the parameterized TCs used here represents distance in units of nautical miles (nmi) and wind speed in knots (kt), with 1 nmi = ~1.15 miles = ~1.85 km and 1 kt = 1 nmi∙h⁻¹ = ~0.51 m∙s⁻¹. Because TC wind radii are typically defined using wind speeds in units of kts, we use nmi and kt in much of the text.

2.1. ECMWF-Derived Parameterized TC Ensemble Forecasts

The operational ECMWF ensemble uses a global numerical weather prediction model to forecast a range of possible future states given initial condition and model uncertainties [62]. We obtained ECMWF data from the TIGGE archive [63,64] at five initialization times: 0 UTC and 12 UTC 8 October 2018, 0 UTC and 12 UTC 9 October 2018, and 0 UTC 10 October 2018, with forecast output at 6 h intervals. The ensemble members were run with an effective grid spacing of ~18 km and were downloaded on a 0.125° latitude × 0.125° longitude grid. The first initialization time is less than 12 h after Michael became a tropical storm, and the last initialization time is approximately 18 h before Michael’s landfall (Table 1).

Although the ECMWF ensemble data provide a useful representation of the meteorological forecast uncertainty associated with Hurricane Michael, they have two limitations for simulating TC-induced storm surge using ADCIRC. First, when the gridded ECMWF data are used to force ADCIRC, the model’s wind and pressure interpolation algorithm does not smoothly translate the storm as a coherent structure between the 6 h meteorological updates. Instead, the storm fades out at the initial location and in at the new location, which induces inaccuracies in the storm surge simulation. Second, at all five initialization times examined here, all of the ECMWF ensemble members forecast TCs with substantially lower peak Vmax than both OFCL and Best track (Table 1, left hand side of Figure 3). This intensity bias occurs for several reasons, including (1) with ~18 km grid spacing, the atmospheric model cannot resolve the hurricane’s inner core and (2) with the surface layer parameterizations used by ECMWF in 2018, the model predicted 10 m wind speeds that were too low in hurricanes [41,65,66,67,68].

To address these limitations, we developed a method for parameterizing the ECMWF ensemble members’ TCs (Section 2.1.1), and then adjusting the TCs’ Vmax using OFCL (Section 2.1.2). NHC issues updated OFCL forecasts at 3, 9, 15, and 21 UTC, and the operational ECMWF ensemble is available several hours after its initialization time. Thus, to correspond to each ECWMF initialization time, we chose the OFCL issued 3 h later. After scaling each parameterized TC’s Vmax, we adjusted its Pmin to reflect the new Vmax (Section 2.1.2) and then estimated missing wind radii (Section 2.1.3).

Although the TCs in OFCL reach higher intensities than those in the ECMWF ensemble, OFCL still substantially underpredicts Hurricane Michael’s peak Vmax, especially at longer lead times (Table 1). OFCL also only provides forecast data every 12 or 24 h (Figure 3), which limits the temporal resolution of the intensity scaling. However, rather than scaling TC intensity using Best track or another source, we used OFCL to simulate forecast information available at the time each ECMWF ensemble forecast was available.

2.1.1. Parameterizing ECMWF Ensemble Members

To derive TC parameters for each ECMWF ensemble member and forecast time, we first obtained TC center locations from the ECMWF A-deck files. We then used the gridded ECMWF data to identify the maximum 10 m wind speed within 250 km of each TC center (Vmax) and minimum sea level pressure within 100 km (Pmin). We also calculated the maximum distance of 34 kt, 50 kt, and 64 kt winds within 300 nmi (555.6 km) of each TC center (R34, R50, and R64) in each of the four quadrants. This process generated, for each initialization time and ensemble member, a sequence of parameterized TCs at 6 h intervals that represents the storm’s evolution in the ECMWF forecast data.

2.1.2. Adjusting TC Intensity

To adjust TC intensity for each initialization time, we first scaled the evolution of mean Vmax10 in the parameterized TC ensemble using the evolution of Vmax10 in the corresponding OFCL, as described in further detail in Appendix A. We then estimated the Pmin of each parameterized TC in the ensemble based on its new Vmax, using the wind-pressure relationship in [69] as described in Appendix B. Together, these two steps generate ensembles of parameterized TC forecasts whose tracks match the TC tracks in the ECMWF ensemble members, but with reduced intensity bias.

Note that this is only a first-order approach to adjusting forecast TC intensity in atmospheric model output, which could be improved upon in future work. Compared to approaches that adjust TC intensity and structure by blending parameterized vortices with atmospheric model data [5,66], using only parameterized vortices limits our ability to represent winds away from the TC core, but it facilitates using a variety of types of already-available atmospheric model data.

2.1.3. Estimating Missing Wind Radii

For some TCs, the intensity adjustment described in Section 2.1.2 increased Vmax from below 34, 50, or 64 kts to above those thresholds. When this occurs, R34, R50, and/or R64 values are missing for that ensemble member and forecast time. Preliminary simulations indicated that, as illustrated in Section 3.2, missing wind radii can substantially affect a parameterized TC’s structure and the storm surge it generates in ADCIRC. To address this, we developed and implemented a method for filling in missing wind radii in the intensity-adjusted TCs, using available wind radii for that TC ensemble member and information derived from the wind radii in the corresponding OFCL forecast. Details of this method are presented in Appendix C.

We developed this method iteratively, by testing it for a variety of ensemble members, examining the resulting ADCIRC simulations, and then making adjustments to address cases of unrealistic TC wind fields or storm surge. Like the TC intensity adjustment method, this approach could be improved upon in future work. However, the method is simple, implementable in real time, and, as shown in Section 3.2, effective for the purpose of this study.

2.2. Storm Surge and Coastal Inundation Simulations

To simulate storm surge and coastal inundation, we ran with ADCIRC in two-dimensional mode, using the same setup as [16,17] with an updated version of the ADCIRC code. ADCIRC uses a finite-element, variable resolution mesh; here, we use a mesh developed and validated by Riverside Technology and AECOM with the National Ocean Service [70,71]. This mesh was developed for operational use by NOAA and has been used in a number of research studies (e.g., [16,17,19,27,72,73]). It contains approximately 1.8 million nodes spanning the Gulf of Mexico and U.S. Atlantic coast and inland to 10 m elevation. The node spacing is coarse in the open ocean and decreases to less than 200 m near the coast (Figure 4), which enables it to represent complex coastal shorelines, bathymetry, and topography at that resolution.

Simulations were initiated with a cold start run beginning at 0 UTC 22 September 2018, without meteorological forcing. Astronomical tides were ramped up to full force over six days, and then the cold start run was continued for another nine days to allow the basin to come into tidal equilibrium without transient gravity waves. Conditions were output from the cold start run at 0 UTC 7 October 2018 and used to initiate the Best track simulations, and subsequently the ensemble forecasts, as described below. Tidal forcing is continued throughout all of the simulations.

2.2.1. Best Track Simulation and Best Track Experiments

To generate an estimated “truth” simulation, we used ADCIRC to simulate the storm surge and coastal inundation associated with Hurricane Michael’s Best track. As shown in Figure 3, Michael’s Best track includes data at 6 h intervals, with one additional data point around the time of peak intensity. To run the Best track simulation, we used output from the cold start run to hot-start ADCIRC at 0 UTC 7 October with Best track forcing added and ramped up to full strength over 1 day (using a hyperbolic tangent ramp function).

We also present results from four Best track experiments, two that systematically investigate the effects of missing wind radii and two that test the method used to estimate missing wind radii. These experiments used the same evolution of TC position, Vmax, and Pmin as the control Best track simulation (BT_control), but with modified wind radii:

• BT_removeR64: R64 (64 kt wind radii) values removed in all 4 quadrants
• BT_removeR64R50: R64 and R50 (64 kt and 50 kt wind radii) values removed in all 4 quadrants
• BT_replaceR64: R64 removed in all 4 quadrants, then estimated as described in Section 2.1.3
• BT_replaceR64R50: R64 and R50 removed in all 4 quadrants, then estimated as described in Section 2.1.3

The first two experiments leave some wind radii unconstrained, so that the GAHM fits a vortex with the available wind radii. For the latter two experiments, we estimated wind radii using R64:R50 and R50:R34 ratios calculated using Best track rather than OFCL. ADCIRC simulations for these experiments were run in the same way as for Best track, with TC wind radii modified throughout the simulation.

2.2.2. TC–Storm Surge Ensemble Forecasts

For the TC–storm surge ensemble simulations, we output conditions from the Best track simulation at the ensemble initialization time, and then hot-started ADCIRC runs forced by the sequence of parameterized TCs in each of the intensity-adjusted TC ensemble members. This generates, for each initialization time, a 50-member ensemble of TC–storm surge forecasts. For the 8 and 9 October initializations, TC forcing was ramped up to full strength over 1 day. For the 10 October initialization, we used a 0.5 day ramp so that the TCs reached full intensity prior to landfall.

We hot-started the ensemble forecasts from the Best track simulation to account for the cumulative effects of the storm up to the initialization time. We also tested an alternate approach for generating storm surge ensembles, by initializing ADCIRC at 0 UTC 7 October with output from the cold start run and prepending Best track on each TC ensemble member. The two approaches generated similar storm surge ensembles, indicating that a hot-start approach with a 1.0 or 0.5 day ramp worked well, at least for Hurricane Michael and the forecast lead times studied here. Thus, we used the hot-start approach to increase computational efficiency.

As described in Section 2.1, all wind speeds (Vmax and wind radii) in the parameterized TC ensemble forecasts represent V10. Thus, the TC–storm surge ensemble forecasts were run with a version of ADCIRC recompiled after setting one2ten to 1.0 in the Fortran code, to remove the conversion from V1 to V10 that ADCIRC typically uses with parameterized TC input [60].

2.3. Data Analysis

To validate that the simulation methodology produces realistic results, we compared storm surge output from the Best track simulation with observations taken during Hurricane Michael in the primary area of interest. Based on [38], we selected the 17 USGS water level sensors (https://stn.wim.usgs.gov/FEV/#2018Michael, accessed on 14 November 2023; [74]) shown in Figure 4 and

Table 2. Locations of the 17 USGS water level sensors used in the study and their maximum observed water levels (m, NAVD88) during Hurricane Michael.

Location Number	Station Name	Station Site Number	Latitude (°N)	Longitude (°E)	Observed Maximum Water Level (m)
1	Shalimar	FLOKA03301	30.4434	−86.5844	1.009
2	Panama City Beach	FLBAY26247	30.1316	−85.7432	1.573
3	Mexico Beach	FLBAY03283	29.9490	−85.4246	4.740
4	Port St. Joe	FLGUL26254	29.7268	−85.3914	2.411
5	Apalachicola	FLFRA03276	29.7232	−84.9830	2.505
6	East Point—St. George Island State Park	FLFRA26257	29.7031	−84.7619	2.478
7	Panacea—Ochlockonee Bay	FLWAL03369	29.9770	−84.3840	2.566
8	Alligator Point	FLFRA26263	29.8939	−84.3736	2.643
9	St. Marks	FLWAK03364	30.1518	−84.2090	2.859
10	Aucilla River	FLTAY17325	30.1165	−83.9795	2.743
11	Ecofina River	FLTAY03362	30.0586	−83.9066	2.667
12	Perry—Spring Warrior Fish Camp	FLTAY03359	29.9201	−83.6704	2.518
13	Keaton Beach	FLTAY03356	29.8189	−83.5949	2.350
14	Dark Island	FLTAY25003	29.8040	−83.5888	2.356
15	Hagens Cove	FLTAY03355	29.7730	−83.5795	2.326
16	Big Bend Wildlife Management Area	FLTAY24950	29.7215	−83.4865	2.225
17	Steinhatchee River	FLDIX03354	29.6701	−83.3891	2.067

. We matched each observation location with the closest wet ADCIRC node (given the mesh spacing of ~200 m near the coast), converted the observation data from feet to meters, and then compared simulated and observed maximum water levels.

To compare the Best track experiments with the control Best track simulation, we examined their parameterized TC wind fields, spatial distribution of the storm surge, and time series of inundation volume. We use inundation volume as a spatially integrated summary measure of coastal inundation; following [17], we calculated inundation volume at a given time by identifying each ADCIRC node that is normally dry (above Mean High High Water, MHHW) and has inundation depth ≥ 1 m, and then summing the inundation volume (inundation depth × cell area) of those nodes. We examine results for inundation ≥ 1 m because this corresponds to the inundation threshold used by the National Hurricane Center (3 feet = 0.91 m) for issuing storm surge watches and warnings, although as described below, we also used other inundation thresholds. To filter out inundation in outlying areas unrelated to TC-induced surge, we calculated inundation volume over a regional domain (92–82° W, 26–31.5° N), excluding the leftmost 2.5% and rightmost 2.5% of nodes in the inundation area [17].

To analyze results from the storm surge ensemble forecasts, we used the control Best track simulation as “truth” to enable comparison across all ADCIRC nodes. Results are shown for three metrics: (1) maps of probability of inundation depth ≥ 1 m (above ground), for ADCIRC nodes above MHHW [16], (2) time series of inundation volume (calculated as described above), and (3) Brier skill score (BSS) [75]. We use the first metric to compare the spatial distribution of predicted coastal inundation at the different initialization times with each other and with the Best track simulation. The second metric serves as a summary measure for exploring how inundation magnitude and timing vary across the 50 ensemble members and how this changes with forecast lead time. We also use this spatially integrated metric to assess large-scale differences in inundation severity across the ensemble and compare the range of ensemble scenarios with Hurricane Michael’s actual inundation. The third metric, BSS, assesses ensemble forecast skill at each initialization time by comparing the predicted probability of inundation greater than or equal to a given threshold (1 m or 0.3 m) to the Best track simulation. BSS was calculated for nodes at or above MHHW within the same domain as the inundation volume (92–82° W, 26–31.5° N). The fraction of nodes within that domain with inundation greater than or equal to that threshold in the Best track simulation was used as the reference forecast. A higher BSS indicates a more skillful forecast, with a maximum of 1.0 for a perfect forecast.

The final set of results analyzes relationships between an ensemble member’s TC parameters and its simulated maximum inundation volume, for the three 0 UTC initialization times. Results are presented for five TC parameters that prior research has found influence storm surge: longitude (representing landfall location), direction (angle), Vmax (intensity), R34 in the TC’s SE quadrant (storm size), and forward speed [17,19,23,24,25,76,77,78,79]. Direction is measured clockwise relative to north = 0°. We used R34 as a measure of storm size because many of the ensemble TCs lacked R50 and R64 values, and we used size in the SE quadrant because for Hurricane Michael it contains most of the onshore winds that are responsible for much of the storm surge [80]. Relationships were examined for each TC parameter at landfall and 6 h before and after landfall [76,77,81].

3. Results: Best Track Simulations

3.1. Best Track Simulation and Comparison with Observations

Figure 5 depicts results from the Best track ADCIRC simulation. The general spatial patterns of simulated storm surge are similar to observations and previous studies of Hurricane Michael, with the highest water levels just to the east of the storm’s track, an extended area of elevated water levels further east along the Florida coast, and a much smaller increase in water levels west of the track [5,30,38,39,82]. The extended area of elevated water levels to the east occurs in the Big Bend region of Florida, where the coastline transitions from approximately east–west along the Florida Panhandle to a more north–south orientation along the Florida Peninsula. This region is highly susceptible to storm surge due to its concave coastline combined with shallow bathymetry along the Florida shelf [83].

Figure 5a and Figure 6 compare maximum water level in the Best track simulation with observations. The correlation (R² = 0.79) is on the high end of other published simulations of Hurricane Michael’s storm surge (range of R² = 0.16–0.74), and the root-mean-square error (RMSE = 0.6 m) is within the range of previous studies (0.26–0.63 m; [5,30,39,82]). The simulated maximum water level at Mexico Beach (location 3), the station with the highest observed water level, is within 10% of the observed value. The observation stations with the largest errors (locations 1–2, 4–6, and 8) were all located on or behind a narrow peninsula or barrier island, where the ADCIRC node spacing may be insufficient to simulate the details of storm surge. Illustrating the importance of such details for accurately simulating water levels at specific locations, if observation location 4 is shifted less than 0.002° (several hundred meters) to the east, the closest ADCIRC node shifts from the ocean side to the bay side of St Joseph peninsula. Because the simulated maximum water level on the ocean side of the peninsula is more than 1 m higher than on the bay side (Figure 5a), this small shift changes the overprediction in Figure 6 to an underprediction. Further east (locations 9–17), simulated maximum water levels have smaller errors.

Figure 6 also indicates that simulated maximum water levels are lower than those observed at most of the gauge locations. One possible explanation is that the simulation does not include wave heights, which [84] indicates were 0.5 m or greater for Michael at several of the observation locations examined here. Another is that the atmospheric forcing used here only includes the simulated TC wind fields, not the large-scale easterly flow that observation time series indicate increased water levels prior to the storm’s arrival. Supporting this latter explanation, ref. [5] found that simulating Hurricane Michael’s storm surge with larger-scale winds added to the GAHM vortex winds increased water levels by 0.1–0.4 m across the region (see also [72]).

In summary, these results indicate that our methods for simulating storm surge are robust, and thus are suitable for running the additional simulations shown in Figure 2. In addition, the Best track simulation is a sufficiently realistic representation of Hurricane Michael’s storm surge that it can be used to evaluate the TC–storm surge ensemble forecasts.

3.2. Sensitivity of Best Track Simulation to Missing and Estimated Wind Radii

Figure 7a–c compares results from the control Best track simulation with those from the two Best track simulations with missing wind radii. When Vmax > 64 and only R34 and R50 values are available (BT_removeR64), the GAHM produces parameterized vortices with a larger inner core and larger area with greater than 64 kt winds, which generates more storm surge (Figure 7b). This generates not only substantially greater inundation volume, but also a longer period of inundation (Figure 8). When Vmax > 64 and only R34 values are available (BT_removeR64R50), these effects are exacerbated (Figure 7c and Figure 8).

These issues with missing wind radii are especially pronounced in Best track due to its high Vmax. However, as described in Section 2, we found similar issues in some OFCL-forced storm surge simulations, due to how the parameterized TC algorithm in ADCIRC interpolates between OFCL’s forecast data points (every 12 or 24 h) when they span landfall. For example, for the 12 UTC 8 October initialization time, the OFCL TC makes landfall between the 12 UTC 10 October and 12 UTC 11 October forecast times, during which Vmax decreases from 105 kts to 45 kts (Figure 3). Since OFCL has only 34 kt wind radii at the latter forecast time, between these two forecast times the GAHM begins producing TCs with larger wind fields, as in the BT_removeR64R50 simulation. That leads to a larger area of onshore winds > 64 kt around the time of landfall, along with a longer period of onshore winds > 64 kt after landfall as the storm moves northeast, both of which contribute to spuriously large storm surge. Before we filled in missing wind radii, we also found such issues in several TC–storm surge ensemble members.

Figure 7d,e depicts results when the wind radii that were removed in the BT_removeR64 and BT_removeR64R50 simulations were replaced using the method described in Section 2.1.3. When only R64 is replaced with an estimated value (BT_replaceR64), the parameterized TC wind fields and resulting storm surge are very similar to those in the control Best track simulation (Figure 7d and Figure 8). Replacing both R64 and R50 with estimates (BT_replaceR64R50) alters the wind field and simulated surge, but these changes are substantially smaller than when missing wind radii are not estimated (Figure 7e and Figure 8). Together, these experiments illustrate that the method used for estimating wind radii is reasonable, and that it alleviates issues that can arise when wind radii are missing in certain situations.

4. Results: TC–Storm Surge Ensemble Forecasts

4.1. Intensity-Adjusted TC Ensemble Forecasts at Multiple Lead Times

Before discussing the TC–storm surge ensemble forecasts, we examine key characteristics of the TC ensemble forecasts derived from the ECMWF ensemble at different initialization times, compared to Best track and NHC’s OFCL forecast. Looking first at storm track, Figure 1 shows that at the two 8 October initialization times, the TC ensemble spread covers much of the central U.S. Gulf coast. With each subsequent initialization time, the spread in storm landfall location narrows, along with the spread in storm direction near landfall. At all five initialization times, the ensemble spread in TC track includes the Best track. However, the majority of ECMWF-derived ensemble TCs track west of Best track, similar to other ensemble predictions of Hurricane Michael [40,85]. OFCL, on the other hand, makes landfall near to or east of Best track for all five initialization times.

Ensemble spread in timing of TC landfall also decreases substantially with forecast lead time (Figure 9). For the 8 October initialization times, forecasts of TC landfall time span approximately 48 h. By the 0 UTC 9 October initialization time, this narrows to approximately 18 h, with two late-landfalling outliers. At these first three initialization times, landfall time and location are correlated: ensemble TCs with more western tracks tend to make landfall earlier and those with more eastern tracks tend to make landfall later. By the latter two initialization times, the TCs in all ensemble members make landfall within a 6–12 h period. Another noticeable aspect of Figure 9 is a bias in TC landfall time in the ensemble forecasts: at all five initialization times, most of the TCs in the ECMWF-derived ensembles make landfall later than Hurricane Michael’s actual landfall. In contrast, the five OFCL TC forecasts, which were generated in real time by NHC forecasters, make landfall within a few hours of Best track.

Ensemble variation in TC intensity is more complex (Figure 1 and Figure 3, Table 1). As forecast lead time decreases, ensemble spread in Vmax at a given time tends to decrease, and peak Vmax tends to increase. However, only a few ensemble TCs reach Hurricane Michael’s actual intensity. In addition, at several initialization times, many of the intensity-adjusted ensemble TCs reach Category 3 or 4 strength over the Gulf of Mexico but then drop to Category 1 or tropical storm strength by landfall. This temporal evolution differs from Best track Vmax, which does not begin decreasing until after landfall. This difference arises in part from how TC intensity evolves in the ECMWF ensemble members, and in part from how we scaled ensemble Vmax to OFCL, as discussed in Appendix A.

These features of the TC ensembles provide important context for interpreting subsequent results from the storm surge ensembles. In the next section, it will be shown that, despite some shortcomings of the meteorological aspects of the ensemble forecasts, they produce a realistic representation of the time-evolving, spatial distribution of probabilities of impactful storm surge in the region at risk.

4.2. Prediction and Predictability of TC Storm Surge at Multiple Lead Times

Next, we analyze how the TC ensemble forecasts discussed in Section 4.1 translate into skill and variability in TC–storm surge ensemble forecasts. Based on Hurricane Michael’s actual evolution, the five initialization times studied here represent lead times of approximately 18–66 h prior to landfall. This is equivalent to approximately 12–60 h before the arrival of Michael’s tropical storm force winds, which is when forecasters and emergency managers recommend having all pre-storm preparations completed.

We start by examining ensemble-derived probabilistic forecasts of inundation depth ≥ 1 m at different locations, shown in Figure 10. At the two 8 October initialization times, the TC–storm surge ensemble indicates that a swath of coastal Florida from the central panhandle to the Big Bend is at risk of substantial inundation (Figure 10a,b). Although many ensemble members’ TCs make landfall west of Best track, in the western Florida panhandle, the storm surge ensemble indicates that this region is at lower risk. Instead, the area at risk identified by the 8 October ensembles extends further east than the eventual ≥ 1 m inundation (Figure 10f), reflecting the Big Bend region’s greater susceptibility to storm surge. At these lead times, locations at risk typically have inundation probabilities between 5 and 45%. However, a few areas have probabilities exceeding 65%, indicating that more than 2 days before landfall, the ensemble identifies these locations as highly likely to experience substantial inundation.

At the two 9 October initialization times, the area at risk of inundation ≥ 1 m narrows (Figure 10c,d), consistent with decreasing spread in the TC ensemble. Moreover, at many of the locations inundated in the Best track simulation, the inundation probabilities predicted by the ensemble increase, exceeding 85% at 0 UTC 9 October and 95% 12 h later. At 0 UTC 10 October, the ensemble further narrows the area at risk, although there remain locations with some probability of inundation ≥ 1 m that do not experience this level of inundation in the Best track simulation. This progression in skill at predicting the probability of inundation ≥ 1 m at specific locations is reflected in the increase in BSS with initialization time, shown in

Table 3. Skill of probabilistic predictions from the TC–storm surge ensemble for each of the five initialization times.

Initialization Time	Approximate Lead Time Prior to Landfall	Brier Skill Score
		Inundation ≥ 1 m	Inundation ≥ 0.3 m
0 UTC 8 October	66 h	0.57	0.72
12 UTC 8 October	54 h	0.57	0.69
0 UTC 9 October	42 h	0.66	0.76
12 UTC 9 October	30 h	0.72	0.77
0 UTC 10 October	18 h	0.82	0.84

. Forecasts of the probability of a lower level of inundation—0.3 m—show a similar progression in skill, but higher BSS, especially for earlier initialization times (Table 3).

Figure 11 depicts how ensemble forecasts of inundation magnitude and timing evolve with lead time. Consistent with the variability in TC track and timing (Figure 1a,b and Figure 3), the two ensembles initialized on 8 October exhibit substantial variability in both maximum inundation volume and inundation timing (Figure 11a,b). They also exhibit a substantial late bias in forecast inundation timing, corresponding to their late bias in forecast timing of TC landfall. With each subsequent initialization time, ensemble variability in inundation timing narrows further (Figure 11c–e). However, even for the 0 UTC 10 October initialization time, inundation volume varies substantially across the ensemble. By this time, the ensemble TCs’ landfall locations vary by only about 80 km, but smaller differences in TC track and other characteristics lead to variability in the magnitude of inundation. Analysis of individual ensemble members revealed that this variability is associated with differences in the extent of flooding along the coast and inland as well as differences in water levels in flooded areas. The contributions of different TC parameters to this variability are investigated further in the next section.

Overall, these results illustrate both the potential for extending storm surge prediction to longer lead times and its challenges. Approximately 2.5–3 days before storm landfall, the TC–storm surge ensemble approach used here is able to identify the major regions at risk of substantial inundation from Hurricane Michael, despite the wide range of TC tracks and the underprediction of even OFCL-scaled Vmax near landfall. As lead time decreases, the TC ensemble indicates decreased meteorological forecast uncertainty, and the storm surge ensemble narrows the area at high risk and likely timing of inundation. Despite these successes, however, even at 18 h before landfall, location-specific predictability of inundation depth is limited [17]. Our results suggest that this occurs because translating the TC ensemble forecasts into storm surge ensemble forecasts attenuates the relationship between decreasing lead time and decreased TC forecast uncertainty.

4.3. Sensitivity of Coastal Inundation to TC Parameters

Qualitative examination of TC–storm surge ensemble members suggested that certain storm characteristics tended to be associated with greater coastal inundation. For example, some of the highest volumes, depths, and areal extents of inundation were produced by TCs that made landfall further east in the Florida Panhandle than Best track and had a more eastward track after landfall, generating onshore winds that lasted for a longer time and extended further east into the susceptible Big Bend region. To explore these associations quantitatively, we extracted different TC parameters from the TC–storm surge ensemble members as described in Section 2.3 and then explored their relationships with maximum inundation volume. However, we anticipate that these relationships depend on the storm and the coastal region affected, as well as the characteristics and limitations of the TC ensemble used. With this in mind, we discuss high-level patterns in the results, with the goal of illustrating how the storm-specific parameterized TC ensemble approach used here can build new understanding about TC-induced storm surge and help elucidate TC–storm surge scenarios.

Figure 12 depicts the relationships between each ensemble member’s maximum inundation volume and its TC’s parameters at landfall for the three 0 UTC initialization times. Consistent with the discussion in Section 4.1, ensemble variability in all five TC parameters tends to decrease as forecast lead time decreases. As discussed in Section 4.2, variability in maximum inundation volume also decreases with lead time, but to a lesser extent. Considering these variables together, the relationships between TC parameters and inundation volume exhibit the most variability for the 0 UTC 8 October initialization. By the 0 UTC 10 October initialization (18 h before landfall), much of the joint variability has collapsed, and several of the relationships appear dominated by noise. In between, 1.5–2-day lead times appear more favorable for considering the relationships between TC parameters and inundation volume; the TC ensemble variability is sufficiently reduced so that relationships emerge for most of the TC parameters, but not so small that inundation is governed by interactions between more detailed aspects of the TC and the coastal geography.

To provide a broader perspective and further illustrate this approach’s potential,

Table 4. Correlations between TC parameters and maximum inundation volume in the TC–storm surge ensemble for the three 0 UTC initialization times. Correlations are shown for the value of five TC parameters 6 h before landfall (−6 h), at landfall (0 h), and 6 h after landfall (+6 h). Italics indicate correlations stronger than ±0.5.

Initialization Time	Approx. Lead Time	Longitude (°E)			Vmax (kt)			R34 in SE Quadrant (nmi)			Direction (°)			Forward Speed (kt)
		−6 h	0 h	+6 h	−6 h	0 h	+6 h	−6 h	0 h	+6 h	−6 h	0 h	+6 h	−6 h	0 h	+6 h
0 UTC 8 October	66 h	0.61	0.59	0.58	0.40	0.59	0.64	0.09	0.18	0.47	0.15	0.07	−0.08	−0.15	−0.01	0.09
0 UTC 9 October	42 h	0.78	0.74	0.72	−0.32	0.47	0.72	0.00	0.19	0.60	0.63	0.52	0.33	−0.68	−0.59	−0.47
0 UTC 10 October	18 h	0.62	0.53	0.41	0.39	0.39	0.42	0.35	0.46	0.59	−0.02	−0.09	−0.29	0.06	−0.03	−0.11

presents correlations for the same results as Figure 12 alongside similar results for the TC parameters 6 h before and after landfall. Longitude exhibits the most consistent relationship with inundation volume, with TCs that track further east tending to produce more inundation across initialization times. Vmax and R34 also exhibit substantial relationships with inundation volume at multiple initialization times, especially for the TC parameters 6 h after landfall. In contrast, direction and forward speed are strongly related to inundation volume in the ensemble initialized at 0 UTC 9 October, but not the earlier or later times. The relationships with Vmax, R34, and forward speed are consistent with prior research indicating that more intense, larger, and slower moving storms tend to produce greater storm surge (e.g., [17,24,25,76,77,78,79]). These results for TC direction differ from some prior studies, which indicate that storms with tracks that make landfall perpendicular to a coastline tend to produce greater surge [24,78]; our results for relationships with direction and longitude are likely due to the position of the storm’s track relative to the Big Bend region.

The ensemble results reveal two additional aspects of Hurricane Michael’s storm surge. First, Figure 12 shows that at each initialization time, multiple ensemble members have TCs that produce substantially higher inundation volume than Best track. This indicates that the flooding caused by Michael’s storm surge could have been much worse, with a slightly different TC track or structure. Moreover, each TC–storm surge ensemble member corresponds to a dynamically and situationally realistic scenario that can be characterized in terms of TC parameters such as those in Table 4. This suggests another potential utility of this study’s ensemble method: highlighting physically realistic atmospheric scenarios that are likely to produce especially severe storm surge impacts. Such analysis could be conducted from a climatological perspective, or to provide information to support decision making in real time during a storm threat [14].

Figure 12 also shows that realistic alternative TC scenarios can produce substantially more inundation than Best track, even with Vmax values up to 70 kts weaker at landfall. This could be related to many of the ensemble TCs reaching peak intensity prior to landfall and subsequently weakening; however, Figure 3 shows that at 0 UTC 8 and 9 October, the highest Vmax that any ensemble TC member reaches is 15 kts weaker than Best track. In other words, although forecasting Hurricane Michael’s rapid intensification was critical for predicting wind impacts, our results suggest that it was not required to provide useful storm surge risk information [86].

5. Discussion

This study developed a method to parameterize and adjust tropical cyclones derived from atmospheric model ensembles, for use in generating storm surge ensembles. Through this approach, we seek to create TC–storm surge ensemble forecasts that capture the atmospheric uncertainty and physical characteristics inherent to an individual TC, while retaining the simplicity of using parametric vortices to simulate storm surge. The article illustrates the potential of this type of ensemble approach for storm surge prediction and predictability research. More generally, it also advances capacity for using global atmospheric model data for storm surge research, forecasting, and risk analyses.

We implemented the method using archived data from the ECMWF ensemble forecasts generated in real time at five initialization times, in conjunction with the ADCIRC dynamical storm surge model. We focus on storm surge forecasts at lead times of approximately one day or longer because that is when most coastal evacuation and preparation decisions are made, and because research indicates that storm surge forecasts are needed more than 48 h in advance for many coastal evacuation decisions [10,11,12,13,14]. Constructing and analyzing TC–storm surge ensemble forecasts across a range of lead times facilitates investigating the extent to which different aspects of forecast skill extend to longer lead times, as well as the TC characteristics that drive forecast variability at different lead times.

We find that the method produces a robust representation of storm surge inundation and associated forecast uncertainty, once parameterized TC intensity is adjusted using NHC’s official forecasts and missing wind radii are estimated. The results illustrate how Hurricane Michael’s storm surge could have been substantially more severe, with a variety of physically realistic TC scenarios represented within the ensembles. Moreover, the ensembles include TCs that have substantially lower intensity than Hurricane Michael at landfall but produce more storm surge inundation. This suggests that predicting Michael’s rapid intensification was not required to predict the storm surge risk posed by the storm—a potentially important result, given the current challenges of accurately predicting rapid hurricane intensification.

The study’s findings suggest that this approach to generating TC–storm surge ensembles has potential to provide forecast utility multiple days before landfall, by identifying areas at substantial risk of storm surge inundation. If such information could reliably be generated in near real time, it could help mobilize early protective actions. The results also illustrate the potential of this approach for understanding which TC parameters are most related to storm surge magnitude in different situations. Note, however, that several of the TC parameters investigated here are correlated in the ensembles, and so a multivariate analysis may be more informative. In addition, despite indicating general patterns, our results show that the relationships between TC parameters and inundation can vary substantially among ensemble members. As discussed in [19], this likely arises due to the complexity of how TCs interact hydrodynamically with local variations in coastline geography, bathymetry, and topography. Consequently, even 18 h before landfall, the storm surge ensemble forecasts exhibit significant spread in the magnitude of inundation. Together, these results highlight the potential for longer lead-time predictions of storm surge risk, as well as the challenges of predicting the details of storm surge inundation even at short lead times.

The method demonstrated here has several potential advantages relative to other methods of generating TC–storm surge ensembles. First, this approach for generating ensemble TC forcing is feasible to run in real time because the parameterized vortices can be generated very rapidly when global ensemble data become available. In addition, because the vortices are parameterized rather than based on gridded data, atmospheric model data available at 6 h intervals are sufficient to run ADCIRC. Further, this approach facilitates highlighting physically realistic scenarios within an ensemble that would produce especially severe impacts in a given TC situation, for impacts measured by maximum water level at specific locations, total inundation volume, or another metric. In these ways, this approach suggests a potential path forward for combining the potential real-time applicability of simpler ensemble storm surge forecasting techniques (e.g., [6,20,22,23]) with the meteorological realism of hindcast surge ensembles that are run with high-resolution atmospheric ensemble simulations (e.g., [15,16]).

The method has several limitations and areas for possible improvement. First, the algorithm that we used to adjust TC intensity in the ECMWF-derived ensemble forecasts does not account for differences in the timing of TC landfall. We chose to use a relatively simple intensity adjustment method as a starting point, to explore methods that could be implemented in near-real time; more complex methods could be investigated in follow-on work. In addition, at four of the five initialization times, the ensembles have a substantial bias in TC landfall time. This bias was not present in NHC’s official forecasts for Hurricane Michael, which suggests the potential for adjusting the parameterized TC ensembles’ timing along with intensity, at least for this storm. Although we did not try to ameliorate the timing bias in this study, doing so would be important for real-time forecasting, to provide reliable forecasts of how much time people at risk have to complete pre-storm preparations. Moreover, in areas with strong tidal variation, even a 6-h change in landfall time can make a large difference in the inundation experienced [15,16,73].

Another potential limitation of this approach is that parameterizing TCs from the atmospheric model simplifies their wind fields. Parameterizing TCs is common in storm surge research and forecasting, but additional work is needed to assess whether doing so as part of this method affects the forecast skill and other properties of the resulting storm surge ensembles. For example, parameterizing the storm’s vortex removed the far-field winds that likely contributed to Hurricane Michael’s inundation along the Florida coast, and the method demonstrated here could be expanded to fully account for the wind forcings that drive storm surge. In addition, for some TCs, adjusting Vmax could produce storms with unrealistically high intensities, and adjusting Vmax and Pmin without adjusting the model-derived wind radii could produce altered or unrealistic TC structures. Moreover, altering forecast parameters such as Vmax and Pmin independently of other parameters such as TC track could degrade the realism of the TC scenarios. To address these issues, a more dynamically informed, multivariate approach to adjusting parameterized TCs may be needed. Finally, the method is limited by the TC scenarios represented in the atmospheric model ensembles used. Since atmospheric model ensembles can be underdispersive and have other biases, it is important to consider an atmospheric ensemble’s characteristics when using the data to create storm surge ensembles and when interpreting the results.

Along with the topics noted above, the article suggests several additional areas for future research. One is implementing the method for additional TCs, to further test the method and investigate the topics addressed here across a wider range of cases. Because the method can use available atmospheric model data at different spatial and temporal resolutions, it can also be extended to other numerical weather prediction ensemble data or a multi-model ensemble. Another potential research topic is investigating how the TC intensity adjustment method used here compares with other methods for addressing TC intensity biases in atmospheric models, when the TC data are used in ensemble storm surge simulations. Also of interest is investigating how storm surge ensembles generated using this approach compare with those generated using systematic perturbations of TC parameters.

From a practical perspective, the approach used here provides a starting point for further work that may help improve operational storm surge ensemble forecasting. For example, because the approach developed here uses ensemble data to specify TC wind fields with a small number of parameters, it can provide a bridge from methods that use climatologically based or ad hoc TC perturbations towards hybrid methods that incorporate information from dynamical atmospheric model ensembles. Implementing this type of approach could enhance the reliability of ensemble storm surge forecasts and extend their utility to longer lead times, supporting protective decision making. More broadly, these methods may be useful for other types of work that involve simulating storm surge using atmospheric model data that do not fully resolve TCs’ structures, has other known biases, or has output available at limited time intervals. This includes storm surge research and risk analyses using archived data from operational weather prediction models, atmospheric reanalyses, and climate simulations.